Deep Q-Learning-Driven Frequency Prism Beamforming With Delay-Adjustable IRS in LEO Satellite Networks

This study proposes an advanced strategy for optimizing multibeam formation and time allocation to meet the growing demand for low-Earth orbit (LEO) satellite communications while reducing communication latency. Traditional multibeam formations, constrained by transmission power limitations, exhibit insufficient scalability. To address this, we introduce a frequency-prism-based multibeam formation approach using delay-adjustable intelligent reflecting surfaces (DA-IRS) equipped with integrated delay elements. By dynamically controlling phase shifts and delays within the DA-IRS, this approach enables efficient resource allocation, effectively minimizing communication latency. The proposed solution uses a control optimization method based on a deep Q-network (DQN). In this framework, the network operation center (NOC), responsible for managing multiple satellites, trains a neural network to model the relationship between control schedules and the associated communication delays. Using this model, the NOC can make decisions that significantly reduce the delay between communication requests and their completion. Simulation results confirm that the proposed method outperforms traditional single-beam control and enhances LEO satellite communication performance. The approach contributes to reducing latency and addressing future communication demands, supporting advancements toward 6G non-terrestrial networks.